Intel announced the performance of the fifth generation Xeon Emerald Rapids

At the SC23 conference, Intel disclosed the internal performance test results of its forthcoming fifth-generation Xeon, Emerald Rapids, as well as performance predictions for the next-generation Xeon, Granite Rapids. These results encompass comparisons with Intel’s fourth-generation Xeon, Sapphire Rapids, and also showcase the HPC benchmarks of Xeon Max processors against AMD’s EPYC Genoa 96-core processors.

The fifth-generation Xeon Emerald Rapids continues to utilize the Eagle Stream platform, maintaining compatibility with the LGA 4677 socket like the current Sapphire Rapids. This new generation Xeon adopts a more optimized design, with a slight increase in core and thread counts, larger cache sizes, and improved performance per watt.

Intel states that Emerald Rapids will offer workload-optimized performance and energy-efficient computing, with memory support upgraded from DDR5-4800 to DDR5-5600, enhancing bandwidth and transfer speeds. The fifth-generation Xeon will also support CXL 1.0/2.0 and up to 80 PCIe 5.0 lanes. Compared to the Sapphire Rapids Xeon 8480 (56-core), the Emerald Rapids Xeon 8592 (64-core) demonstrates a 1.4x increase in AI speech recognition performance, a 1.4x rise in HPC LAMMPS performance, and a 1.2x improvement in media transcoding efficiency.

Anticipated features of the fifth-generation Xeon Emerald Rapids processors include:

  • – Triple the amount of third-level cache
  • – Higher memory speeds
  • – Increased CPU counts (up to 64 cores)
  • – CXL Type 3 high-bandwidth interface
  • – Workload optimization (Intel AE)
  • – Optimized power modes
  • – General performance/functionality improvements up to 17%
  • – Offloading CPU cores using Intel Accelerator Engine for enhanced power efficiency
  • – Significant advancements in inference and training
  • – Built-in AI acceleration with Intel AMX
  • – Ready-to-use optimized software solutions

Additionally, Intel revealed internal performance predictions for the upcoming Granite Rapids Xeon processor, a pure P-Core Xeon, while its pure E-Core counterpart is named Sierra Forest, boasting up to 288 cores. Both will utilize the next-generation Birch Stream platform, featuring two different sockets, LGA 4710 and LGA 7529. Future Granite Rapids will offer more cores, higher frequencies, and the latest Intel AMX with added FP16 support to expand aid for AI developers. The new platform will support 12-channel MCRDIMM memory to address the limitations of memory capacity in larger LLM models.

Performance-wise, compared to the current fourth-generation Xeon Sapphire Rapids, Granite Rapids is expected to achieve a 2.9x increase in AI inference (DeepMD+LAAMPS), a 2.8x increase in memory bandwidth, and a 3x improvement in AI performance. However, these are preliminary estimates, and the final performance may vary.

Finally, Intel showcased some HPC performance benchmarks for the Xeon Max 9480 processor, featuring 56 cores and 64GB of HBM memory. Compared to AMD’s 96-core EPYC 9654, the Xeon Max 9480 shows up to a 30% increase in performance. The TDPs of these processors are 360/350W respectively, with Sapphire Rapids likely having a lower total cost of ownership (TCO) than AMD’s solution. Additionally, AMD’s Genoa-X, featuring a 3D V-Cache, performs better in scenarios limited by memory bandwidth.

Intel plans to launch the fifth-generation Xeon Emerald Rapids on December 14, with the release of Granite Rapids scheduled for the second half of 2024, likely following Sierra Forest’s debut.